pprof is a tool for visualization and analysis of profiling data.
pprof reads a collection of profiling samples in profile.proto format and generates reports to visualize and help analyze the data. It can generate both text and graphical reports (through the use of the dot visualization package).
profile.proto is a protocol buffer that describes a set of callstacks and symbolization information. A common usage is to represent a set of sampled callstacks from statistical profiling. The format is described on the src/proto/profile.proto file. For details on protocol buffers, see https://developers.google.com/protocol-buffers
Profiles can be read from a local file, or over http. Multiple profiles of the same type can be aggregated or compared.
If the profile samples contain machine addresses, pprof can symbolize them through the use of the native binutils tools (addr2line and nm).
pprof operates on data in the profile.proto format. Each profile is a collection of samples, where each sample is associated to a point in a location hierarchy, one or more numeric values, and a set of labels. Often these profiles represents data collected through statistical sampling of a program, so each sample describes a program call stack and a number or weight of samples collected at a location. pprof is agnostic to the profile semantics, so other uses are possible. The interpretation of the reports generated by pprof depends on the semantics defined by the source of the profile.
There are few different ways of using
If a report format is requested on the command line:
pprof <format> [options] source
pprof will generate a report in the specified format and exit. Formats can be either text, or graphical. See below for details about supported formats, options, and sources.
Interactive terminal use
Without a format specifier:
pprof [options] source
pprof will start an interactive shell in which the user can type
help to get online help.
If a host:port is specified on the command line:
pprof -http=[host]:[port] [options] source
pprof will start serving HTTP requests on the specified port. Visit
the HTTP url corresponding to the port (typically
in a browser to see the interface.
The objective of pprof is to generate a report for a profile. The report is generated from a location hierarchy, which is reconstructed from the profile samples. Each location contains two values: flat is the value of the location itself, while cum is the value of the location plus all its descendants. Samples that include a location multiple times (eg for recursive functions) are counted only once per location.
options configure the contents of a report. Each option has a value, which can be boolean, numeric, or strings. While only one format can be specified, most options can be selected independently of each other.
Some common pprof options are:
- -flat [default], -cum: Sort entries based on their flat or cumulative weight respectively, on text reports.
- -functions [default], -filefunctions, -files, -lines, -addresses: Generate the report using the specified granularity.
- -noinlines: Attribute inlined functions to their first out-of-line caller.
For example, a command like
pprof -list foo -noinlines profile.pb.gzcan be used to produce the annotated source listing attributing the metrics in the inlined functions to the out-of-line calling line.
- -nodecount= int: Maximum number of entries in the report. pprof will only print this many entries and will use heuristics to select which entries to trim.
- -focus= regex: Only include samples that include a report entry matching regex.
- -ignore= regex: Do not include samples that include a report entry matching regex.
- -show_from= regex: Do not show entries above the first one that matches regex.
- -show= regex: Only show entries that match regex.
- -hide= regex: Do not show entries that match regex.
Each sample in a profile may include multiple values, representing different
entities associated to the sample. pprof reports include a single sample value,
which by convention is the last one specified in the report. The
option selects which value to use, and can be set to a number (from 0 to the
number of values - 1) or the name of the sample value.
Sample values are numeric values associated to a unit. If pprof can recognize
these units, it will attempt to scale the values to a suitable unit for
unit= option will force the use of a specific unit. For
unit=sec will force any time values to be reported in
seconds. pprof recognizes most common time and memory size units.
Samples in a profile may have tags. These tags have a name and a value; this
value can be either numeric or a string. pprof can select samples from a
profile based on these tags using the
Generally, these options work as follows:
- -tagfocus=regex or -tagfocus=range: Restrict to samples with tags matched by regexp or in range.
- -tagignore=regex or -tagignore=range: Discard samples with tags matched by regexp or in range.
-tagignore=regex, the regex will be compared
to each value associated with each tag. If one specifies a value
regex1,regex2, then only samples with a tag value matching
and a tag value matching
regex2 will be kept.
In addition to being able to filter on tag values, one can specify the name of
the tag which a certain value must be associated with using the notation
-tagfocus=tagName=value. Here, the
tagName must match the tag's name
exactly, and the value can be either a regex or a range. If one specifies
a value like
regex1,regex2, then samples with a tag value (paired with the
specified tag name) matching either
regex1 or matching
regex2 will match.
Here are examples explaining how
tagfocus can be used:
-tagfocus 128kb:512kbaccepts a sample iff it has any numeric tag with memory value in the specified range.
-tagfocus mytag=128kb:512kbaccepts a sample iff it has a numeric tag
mytagwith memory value in the specified range. There isn't a way to say
-tagfocus mytag=128kb:512kb,mytag2=128kb:512kb. Just single value or range for numeric tags.
-tagfocus someregexaccepts a sample iff it has any string tag with
tagName:tagValuestring matching specified regexp. In the future, this will change to accept sample iff it has any string tag with
tagValuestring matching specified regexp.
-tagfocus mytag=myvalue1,myvalue2matches if either of the two tag values are present.
-tagignore works similarly, except that it discards matching samples, instead
of keeping them.
If both the
-tagfocus expressions (either a regexp or a
range) match a given sample, then the sample will be discarded.
pprof text reports show the location hierarchy in text format.
- -text: Prints the location entries, one per line, including the flat and cum values.
- -tree: Prints each location entry with its predecessors and successors.
- -peek= regex: Print the location entry with all its predecessors and successors, without trimming any entries.
- -traces: Prints each sample with a location per line.
pprof can generate graphical reports on the DOT format, and convert them to multiple formats using the graphviz package.
These reports represent the location hierarchy as a graph, with a report entry represented as a node. Solid edges represent a direct connection between entries, while dotted edges represent a connection where some intermediate nodes have been removed. Nodes are removed using heuristics to limit the size of the graph, controlled by the nodecount option.
The size of each node represents the flat weight of the node, and the width of each edge represents the cumulative weight of all samples going through it. Nodes are colored according to their cumulative weight, highlighting the paths with the highest cum weight.
- -dot: Generates a report in .dot format. All other formats are generated from this one.
- -svg: Generates a report in SVG format.
- -web: Generates a report in SVG format on a temp file, and starts a web browser to view it.
- -png, -jpg, -gif, -pdf: Generates a report in these formats,
pprof can also generate reports of annotated source with samples associated to them. For these, the source or binaries must be locally available, and the profile must contain data with the appropriate level of detail.
pprof will look for source files on its current working directory and all its
ancestors. pprof will look for binaries on the directories specified in the
$PPROF_BINARY_PATH environment variable, by default
%USERPROFILE%\pprof\binaries on Windows). It will look binaries up by name,
and if the profile includes linker build ids, it will also search for them in
a directory named as the build id.
pprof uses the binutils tools to examine and disassemble the binaries. By
default it will search for those tools in the current path, but it can also
search for them in a directory pointed to by the environment variable
- -list= regex: Generates an annotated source listing for functions matching regex, with flat/cum weights for each source line.
- -disasm= regex: Generates an annotated disassembly listing for functions matching regex.
- -weblist= regex: Generates a source/assembly combined annotated listing for functions matching regex, and starts a web browser to display it.
pprof can subtract one profile from another, provided the profiles are of compatible types (i.e. two heap profiles). pprof has two options which can be used to specify the filename or URL for a profile to be subtracted from the source profile:
-diff_base= profile: useful for comparing two profiles. Percentages in the output are relative to the total of samples in the diff base profile.
-base= profile: useful for subtracting a cumulative profile, like a golang block profile, from another cumulative profile collected from the same program at a later time. When comparing cumulative profiles collected on the same program, percentages in the output are relative to the difference between the total for the source profile and the total for the base profile.
The -normalize flag can be used when a base profile is specified with either
-diff_base or the
-base option. This flag scales the source profile so
that the total of samples in the source profile is equal to the total of samples
in the base profile prior to subtracting the base profile from the source
profile. Useful for determining the relative differences between profiles, for
example, which profile has a larger percentage of CPU time used in a particular
When using the -diff_base option, some report entries may have negative values. If the merged profile is output as a protocol buffer, all samples in the diff base profile will have a label with the key "pprof::base" and a value of "true". If pprof is then used to look at the merged profile, it will behave as if separate source and base profiles were passed in.
When using the -base option to subtract one cumulative profile from another collected on the same program at a later time, percentages will be relative to the difference between the total for the source profile and the total for the base profile, and all values will be positive. In the general case, some report entries may have negative values and percentages will be relative to the total of the absolute value of all samples when aggregated at the address level.
pprof can read profiles from a file or directly from a URL over http or https. Its native format is a gzipped profile.proto file, but it can also accept some legacy formats generated by gperftools.
When fetching from a URL handler, pprof accepts options to indicate how much to wait for the profile.
- -seconds= int: Makes pprof request for a profile with the specified duration in seconds. Only makes sense for profiles based on elapsed time, such as CPU profiles.
- -timeout= int: Makes pprof wait for the specified timeout when retrieving a profile over http. If not specified, pprof will use heuristics to determine a reasonable timeout.
pprof also accepts options which allow a user to specify TLS certificates to use when fetching or symbolizing a profile from a protected endpoint. For more information about generating these certificates, see https://docs.docker.com/engine/security/https/.
- -tls_cert= /path/to/cert: File containing the TLS client certificate to be used when fetching and symbolizing profiles.
- -tls_key= /path/to/key: File containing the TLS private key to be used when fetching and symbolizing profiles.
- -tls_ca= /path/to/ca: File containing the certificate authority to be used when fetching and symbolizing profiles.
pprof also supports skipping verification of the server's certificate chain and host name when collecting or symbolizing a profile. To skip this verification, use "https+insecure" in place of "https" in the URL.
If multiple profiles are specified, pprof will fetch them all and merge them. This is useful to combine profiles from multiple processes of a distributed job. The profiles may be from different programs but must be compatible (for example, CPU profiles cannot be combined with heap profiles).
pprof can add symbol information to a profile that was collected only with address information. This is useful for profiles for compiled languages, where it may not be easy or even possible for the profile source to include function names or source coordinates.
pprof can extract the symbol information locally by examining the binaries using the binutils tools, or it can ask running jobs that provide a symbolization interface.
pprof will attempt symbolizing profiles by default, and its
provides some control over symbolization:
-symbolize=none: Disables any symbolization from pprof.
-symbolize=local: Only attempts symbolizing the profile from local binaries using the binutils tools.
-symbolize=remote: Only attempts to symbolize running jobs by contacting their symbolization handler.
For local symbolization, pprof will look for the binaries on the paths specified
by the profile, and then it will search for them on the path specified by the
$PPROF_BINARY_PATH. Also, the name of the main binary can
be passed directly to pprof as its first parameter, to override the name or
location of the main binary of the profile, like this:
pprof /path/to/binary profile.pb.gz
By default pprof will attempt to demangle and simplify C++ names, to provide
readable names for C++ symbols. It will aggressively discard template and
function parameters. This can be controlled with the
option. Note that for remote symbolization mangled names may not be provided by
the symbolization handler.
-symbolize=demangle=none: Do not perform any demangling. Show mangled names if available.
-symbolize=demangle=full: Demangle, but do not perform any simplification. Show full demangled names if available.
-symbolize=demangle=templates: Demangle, and trim function parameters, but not template parameters.